Insilico Medicine Highlights AI Benchmark Results for Drug Potency Prediction

A LinkedIn post from Insilico Medicine highlights results from Day 12 of its ScienceAIBench series, focusing on AI prediction of drug bioactivity via IC50 benchmarks. The exercise evaluates how well leading foundation models predict the binding strength of drug candidates to biological targets using the BindingDB Cold Drug split from the Therapeutics Data Commons.

According to the post, the benchmark measures performance using Spearman and Pearson correlations on IC50 prediction, emphasizing generalization to unseen chemical scaffolds. Opus 4.5 is reported as the top performer, with Spearman and Pearson correlations of 0.347 and 0.349, respectively, while GPT 5.2 follows with more stable behavior and fewer extreme outliers.

The post also notes that Deepseek 3.2 showed the weakest performance, with a Spearman correlation of 0.128, suggesting more limited ability to handle novel chemical space under these conditions. All evaluated models reportedly produced negative R² values, underscoring the difficulty of accurately modeling complex structure–activity relationships in a fully out-of-distribution drug discovery setting.

For investors, the benchmark series suggests Insilico Medicine is positioning itself as an authority on AI model evaluation in small-molecule drug discovery, which could enhance its visibility with pharma and biotech partners. Demonstrating rigorous, task-specific benchmarks in areas such as IC50 prediction may support the company’s value proposition for AI-driven drug design platforms and inform future collaborations, licensing opportunities, or data partnerships.

At the same time, the moderate correlation levels and negative R² reported across models indicate that current general-purpose AI systems still face substantial limitations in high-stakes medicinal chemistry tasks. This gap may highlight ongoing demand for domain-specialized models and proprietary data assets, potentially favoring companies like Insilico Medicine that can build or integrate tailored solutions and justify sustained R&D investment in this niche.